Traditional applications are built around deterministic services, APIs and predefined workflows. Agentic applications introduce something fundamentally different: components that can reason, choose tools, maintain context and dynamically decide what to do next.

So what happens to software architecture when an AI agent becomes an active participant in the system?

This session explores architectural patterns for agentic systems, including single-agent versus multi-agent designs, orchestration, tool boundaries, state and memory, event-driven workflows, human-in-the-loop controls and failure isolation.

We'll use real-world engineering scenarios to examine where agents fit naturally into existing architectures, where traditional deterministic services remain the better choice, and how to avoid turning every application into an unnecessarily complex multi-agent system.

The goal is to provide architects and experienced engineers with a practical framework for deciding when to use an agent, where to place it, and how to engineer the boundaries around it.